Machine learning is a subset of artificial intelligence. Artificial intelligence employs sophisticated algorithms to make decisions similar to those a person might make. Generally, these algorithms are programmed by humans who instruct the computer “how to think” using computer code. Machine leaning advances this technology. In machine learning, patterns in data are detected and used to improve performance based on feedback. These program adjustments are performed by the computer itself without explicit programing by a human.
Up until recently, it has mainly been confined to academic research. There have been few practical applications for machine learning, but one of these is the self-driving car. This type of car uses machine learning to assess driving conditions and modify driving style. In the future, it may also be used to evaluate the health of the driver and re-route to a hospital if necessary.
Machine learning also has great potential in biomedical applications. IBM’s Watson will use computer vision to detect patterns in 30 billion medical images. The goal is to use these data to help doctors diagnose diseases faster and more accurately. Other potential applications of machine learning in medicine include developing personalized medicine with treatments designed specifically for individuals.
Machine learning can be categorized into two general classes. The first is supervised learning, which uses a training data set and solutions to train the algorithm. Modifications are made to the machine-learning data base to minimize errors between predicted output and the known solutions.
The other category is unsupervised learning, where the algorithm tries to find and develop relationships that detect patterns in data, subdivide data sets, or categorize data. Learning is performed without known solutions so it can be open to the discovery of previously unknown concepts or ideas.
It is also possible for machine learning methods to share properties of unsupervised and supervised learning. Reinforcement algorithms do not provide the solution immediately, as in supervised learning, but provide it after a time-delay. This allows the machine to learn on its own first and then begin to develop a possibly unique approach. The initial idea is then refined using actual solutions.
